Tars

[New] We are developing a PyTorch version of Tars made with more sophisticated API (which is named Pixyz). It is released now!

Tars is a deep generative models library. It has the following features:

• Various distributions

  • Gaussian, Bernoulli, Laplace, Gamma, Beta, Dirichlet, Bernoulli, Categorical, and so on.
  • Hierarchical latent distributions (New!).
  • We can draw samples from these distributions by the reparameterization trick .

• Various models

  • Autoencoder
  • VAE
    • Conditional VAE
    • Importance weighted autoencoder
    • JMVAE
    • Multiple latent layers
  • GAN, Conditional GAN
  • VAE-GAN, conditional VAE-GAN
    • JMVAE-GAN
  • VAE-RNN
    • Variational RNN
    • DRAW, Convolutional DRAW

• Various lower bounds

  • The evidence lower bound (ELBO, which is the same as the original lower bound)
  • The importance sampling lower bound
  • The variational R'enyi bound

• Note: Some of the implementations of the above models have not yet been released in this version. If you want to use such models, please use the old version (v0.0.2).

• For a more detailed explanation of this library, please refer to this page (in Japanese).

Installation

$ git clone https://github.com/masa-su/Tars.git
$ pip install -e Tars --process-dependency-links

or

$ pip install -e git://github.com/masa-su/Tars --process-dependency-links

When you execute this command, the following packages will be automatically installed in your environment:

• Theano
• Lasagne
• progressbar2
• matplotlib
• sklearn

Examples

Please go to the examples directory and try to run some examples.